Gryff - Unifying Consensus and Shared Registers

Explore a 20-minute conference talk from USENIX NSDI '20 that introduces Gryff, a novel system unifying consensus and shared registers to achieve linearizability and low tail latency in geo-replicated storage. Delve into the design, implementation, and evaluation of Gryff, which employs carstamps to efficiently order reads and writes without unnecessary constraints. Learn how this approach optimizes shared register protocols in combination with EPaxos, resulting in lower service-level latency compared to EPaxos or MultiPaxos, particularly for read operations. Gain insights into the tradeoffs between linearizability and low tail latency in distributed systems, and understand how Gryff addresses the inefficiencies of traditional consensus-based approaches for workloads dominated by reads and writes.